@ebb-ai/mcp

Model Context Protocol (MCP) server for ebb-ai — exposes carbon-aware task scheduling as MCP tools that any MCP-compatible agent can call.

Install (development, monorepo)

# from the repo root
pnpm install
pnpm --filter @ebb-ai/mcp build

Run

node packages/mcp-server/dist/server.js

The server speaks MCP over stdio. To wire it into an agent client:

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json on macOS (Linux/Windows have similar paths):

{
  "mcpServers": {
    "ebb-ai": {
      "command": "node",
      "args": [
        "/ABSOLUTE/PATH/TO/ebb-ai/packages/mcp-server/dist/server.js"
      ],
      "env": {
        "EBB_ELECTRICITY_MAPS_API_KEY": "optional - mock data without it",
        "EBB_DEFAULT_REGION": "US-CAL-CISO"
      }
    }
  }
}

Restart Claude Desktop after editing.

Claude Code

Add ebb-ai to your ~/.claude/mcp.json (or workspace .claude/mcp.json):

{
  "mcpServers": {
    "ebb-ai": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/ebb-ai/packages/mcp-server/dist/server.js"]
    }
  }
}

OpenClaw

Add to ~/.openclaw/mcp.json:

{
  "mcpServers": {
    "ebb-ai": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/ebb-ai/packages/mcp-server/dist/server.js"]
    }
  }
}

Tools exposed

| Tool | Purpose | |---|---| | get_grid_forecast(region, hours?) | Hour-by-hour carbon intensity for a region; useful before deciding whether to defer. | | schedule_task(prompt, deadline, model?, region?, carbon_budget_g?) | Queue a task to run in the cleanest window inside the deadline. Returns a task_id. | | check_queue_status(task_id?) | Inspect the queue or one specific task; includes a carbon receipt once the task has completed. |

Environment variables

| Var | Purpose | |---|---| | EBB_ELECTRICITY_MAPS_API_KEY | API key from electricitymaps.com. If unset, the server falls back to deterministic mock data so you can still run the whole stack locally. | | EBB_DEFAULT_REGION | Electricity Maps zone code, default US-CAL-CISO. |

v0.1 limitations

• The server schedules dispatch time, but does not yet call the underlying LLM (the agent is expected to execute the prompt itself once the window arrives). Provider adapters and Batch API integration land in v0.2.
• Queue is in-memory; restarts lose state. SQLite persistence lands in v0.2.
• Single-region grid feed. Multi-region routing lands in v0.3.

See ../../ROADMAP.md for the full roadmap.